基于半监督典型相关分析的多视图维数约简
发布时间：2018-08-02 15:35
【摘要】：为了有效地在半监督多视图情景下进行维数约简,提出了使用非负低秩图进行标签传播的半监督典型相关分析方法。非负低秩图捕获的全局线性近邻可以利用直接邻居和间接可达邻居的信息维持全局簇结构,同时低秩的性质可以保持图的压缩表示。当无标签样本通过标签传播算法获得估计的标签信息后,在每个视图上构建软标签矩阵和概率类内散度矩阵,然后通过最大化不同视图同类样本间相关性的同时最小化每个视图低维特征空间类内变化来提升特征鉴别能力。实验结果表明,所提方法比已有相关方法能够取得更好的识别性能且更鲁棒,是有效的多视图维数约简方法。
[Abstract]:In order to reduce dimensionality effectively in semi-supervised multi-view scenarios, a semi-supervised canonical correlation analysis method using non-negative low-rank graphs for label propagation is proposed. The global linear nearest neighbor captured by non-negative low rank graph can maintain the global cluster structure using the information of direct neighbor and indirect reachable neighbor, and the property of low rank can maintain the compressed representation of graph. When untagged samples obtain the estimated tag information through tag propagation algorithm, soft tag matrix and probabilistic intra-class divergence matrix are constructed on each view. Then the ability of feature identification is improved by maximizing the correlation between the same samples of different views and minimizing the intra-class changes in each view's low-dimensional feature space. The experimental results show that the proposed method can achieve better recognition performance and more robust than the existing methods, and it is an effective multi-view dimension reduction method.
【作者单位】： 九江学院信息科学与技术学院;南京邮电大学自动化学院;
【基金】：国家自然科学基金资助项目(61462048) 九江学院科研项目(2014KJYB019,2014KJYB030,2015LGYB26) 江西省教育厅科学技术研究项目(GJJ151076)
【分类号】：TP391.41
，
本文编号：2159892
[Abstract]:In order to reduce dimensionality effectively in semi-supervised multi-view scenarios, a semi-supervised canonical correlation analysis method using non-negative low-rank graphs for label propagation is proposed. The global linear nearest neighbor captured by non-negative low rank graph can maintain the global cluster structure using the information of direct neighbor and indirect reachable neighbor, and the property of low rank can maintain the compressed representation of graph. When untagged samples obtain the estimated tag information through tag propagation algorithm, soft tag matrix and probabilistic intra-class divergence matrix are constructed on each view. Then the ability of feature identification is improved by maximizing the correlation between the same samples of different views and minimizing the intra-class changes in each view's low-dimensional feature space. The experimental results show that the proposed method can achieve better recognition performance and more robust than the existing methods, and it is an effective multi-view dimension reduction method.
【作者单位】： 九江学院信息科学与技术学院;南京邮电大学自动化学院;
【基金】：国家自然科学基金资助项目(61462048) 九江学院科研项目(2014KJYB019,2014KJYB030,2015LGYB26) 江西省教育厅科学技术研究项目(GJJ151076)
【分类号】：TP391.41
，
本文编号：2159892
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2159892.html
